Help Needed: Training Locally Before Deploying on SpiNNaker (World Model + RL)

[Rohit]

Hi everyone,

I'm Rohit, currently working on a reinforcement learning model that incorporates a world model, and I'm trying to deploy it on SpiNNaker using the Jupyter/cloud interface.

However, I'm facing a major bottleneck my model is relatively complex, and running it directly on SpiNNaker is proving to be very slow, which is making development and debugging quite difficult.

I was wondering if there's a recommended way to train and debug the model locally (perhaps using a simulator that closely mimics SpiNNaker behavior), and then deploy the trained model to SpiNNaker just for inference or final validation?

I’ve come across some online resources, but I’m finding it hard to understand what actually works and what’s compatible with my setup. So I thought I’d reach out to the community for advice or suggestions.

Thank you in advance for your support,

Rohit

[Andrew Rowley]

Hi,

Apologies for not getting back sooner - it has been holiday season here!

Unfortunately the loading of data on to SpiNNaker can be a bit slow.  Can I check if you are using the EBRAINS Jupyter (i.e. lab.ebrains.eu) or the SpiNNaker local one (i.e. sands.cs.man.ac.uk) for this?  If you are using the EBRAINS one, you could try the SpiNNaker local one, as that has more direct access to the machine which might be a bit quicker.  That said, if it is very big, it might not help a lot.

Another question is on how you are doing the training.  Is this done by loading everything up and then running for a while, with training happening during the execution, or is it done by stopping and changing the simulation?  The reason I ask is that once everything is loaded, it generally doesn't take long to run things (normally it is in real-time or near real-time), but if you stop and start, there will of course be more delays.  If I know more about the model you are using for training, and you are stopping and starting, there might be ways to convert to an execution where you just run and train on the machine.

Alternatively, the models on SpiNNaker are generally quite simple, so replicating them should be quite easy.  This will depend on what models you are using though, and also noting that the execution using fixed-point arithmetic, which might need to also be replicated to ensure the same execution on and off the machine.

Let us know if you need more help,

Andrew 🙂

---

From: spinnak...@googlegroups.com <spinnak...@googlegroups.com> on behalf of Rohit <saketr...@gmail.com>
Sent: 26 July 2025 16:46
To: SpiNNaker Users Group <spinnak...@googlegroups.com>
Subject: [SpiNNaker Mailing List] Help Needed: Training Locally Before Deploying on SpiNNaker (World Model + RL)

 

--
You received this message because you are subscribed to the Google Groups "SpiNNaker Users Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to spinnakeruser...@googlegroups.com.
To view this discussion, visit https://groups.google.com/d/msgid/spinnakerusers/da1c78cd-d488-4232-bbb2-205eabf61f84n%40googlegroups.com [groups.google.com].

[Christian Brenninkmeijer]

Hi, 

Another thing to add to what Andrew said is that how the model is setup can greatly affect loading time so it would be helpful for him/us to see your model.

Some simple examples include.

• Avoid using PopulationViews to split large Populations into subsections that do different jobs.
• It will nearly always be better to use multiple Populations
• SPyNNaker will partition large Population so there is NO benefits of scale for larger Populations
• Avoid FromList Connectors
• They require loading the whole list
• Many of the other Connectors like the OneToOne, Fixed number ext can be Generated on Machine
• Use PyNNs ability to generate Random numbers rather than doing the Random In Python and inputting Lists into PyNN
• sPyNNaker can often generate the List on Machine,

Christian

University of Manchester

---

From: spinnak...@googlegroups.com <spinnak...@googlegroups.com> on behalf of Andrew Rowley <Andrew...@manchester.ac.uk>
Sent: 19 August 2025 12:09
To: Rohit <saketr...@gmail.com>; SpiNNaker Users Group <spinnak...@googlegroups.com>
Subject: Re: [SpiNNaker Mailing List] Help Needed: Training Locally Before Deploying on SpiNNaker (World Model + RL)

 

To view this discussion, visit https://groups.google.com/d/msgid/spinnakerusers/CWLP265MB5676572967E976E55BFDCDBFC530A%40CWLP265MB5676.GBRP265.PROD.OUTLOOK.COM [groups.google.com].